Clickbait Spoiling (SemEval 2023 Task)

This project addresses the two subtasks from the SemEval-2023 Clickbait Spoiling Challenge:

• Task 1: Spoiler Type Classification - Classify clickbait posts into three spoiler types: phrase, passage, or multi
• Task 2: Spoiler Generation - Generate spoiler text that satisfies curiosity induced by clickbait posts

Various models and training strategies were tested to effectively tackle both classification and generation tasks.

---

Task 1: Spoiler Type Classification

The approach began with simple models (logistic regression) and gradually transitioned toward more powerful Transformer-based architectures.

Final model:

• Ensemble of microsoft/deberta-v3-large and google/electra-large-discriminator
• Achieved F1 score = 0.76

---

Task 2: Spoiler Generation

Encoder-decoder transformer models were used for generating spoilers from clickbait posts and linked articles.

Final model:

• Fine-tuned t5-base
• Achieved METEOR score = 0.431

Note: Only the best model for both tasks are provided in this repo.

---

Dataset

The official JSONL files provided by the challenge organizers are used for this project:

train.jsonl val.jsonl test.jsonl